def train(self, batch: EpisodeBatch, t_env: int, episode_num: int): # Get the relevant quantities bs = batch.batch_size max_t = batch.max_seq_length rewards = batch["reward"][:, :-1] actions = batch["actions"][:, :] terminated = batch["terminated"][:, :-1].float() mask = batch["filled"][:, :-1].float() mask[:, 1:] = mask[:, 1:] * (1 - terminated[:, :-1]) avail_actions = batch["avail_actions"][:, :-1] # Calculate action policy distribution and entropy mac_out = [] mac_out_entropy = [] self.mac.init_hidden(batch.batch_size) for t in range(batch.max_seq_length - 1): agent_outs = self.mac.forward(batch, t=t, return_logits=True) agent_entropy = multinomial_entropy(agent_outs).mean(dim=-1, keepdim=True) agent_probs = th.nn.functional.softmax(agent_outs, dim=-1) mac_out.append(agent_probs) mac_out_entropy.append(agent_entropy) mac_out = th.stack(mac_out, dim=1) # Concat over time mac_out_entropy = th.stack(mac_out_entropy, dim=1) # Mask out unavailable actions, renormalise (as in action selection) mac_out[avail_actions == 0] = 0 mac_out = mac_out / mac_out.sum(dim=-1, keepdim=True) mac_out[avail_actions == 0] = 0 # Mix action probability and state to estimate joint Q-value mix_loss = self.critic(mac_out, batch["state"][:, :-1]) mask = mask.expand_as(mix_loss) entropy_mask = copy.deepcopy(mask) mix_loss = (mix_loss * mask).sum() / mask.sum() # Adaptive Entropy Regularization entropy_loss = (mac_out_entropy * entropy_mask).sum() / entropy_mask.sum() entropy_ratio = self.entropy_coef / entropy_loss.item() mix_loss = -mix_loss - entropy_ratio * entropy_loss # Optimise agents self.agent_optimiser.zero_grad() mix_loss.backward() grad_norm = th.nn.utils.clip_grad_norm_(self.agent_params, self.args.grad_norm_clip) self.agent_optimiser.step() if t_env - self.log_stats_t_agent >= self.args.learner_log_interval: self.logger.log_stat("mix_loss", mix_loss.item(), t_env) self.logger.log_stat("entropy", entropy_loss.item(), t_env) self.logger.log_stat("agent_grad_norm", grad_norm, t_env) self.log_stats_t_agent = t_env
def train(self, batch: EpisodeBatch, t_env: int, episode_num: int): self.train_critic_td(batch, t_env, episode_num) # Get the relevant quantities bs = batch.batch_size max_t = batch.max_seq_length rewards = batch["reward"][:, :-1] actions = batch["actions"][:, :] terminated = batch["terminated"][:, :-1].float() mask = batch["filled"][:, :-1].float() mask[:, 1:] = mask[:, 1:] * (1 - terminated[:, :-1]) avail_actions = batch["avail_actions"][:, :-1] mac_out = [] mac_out_entropy = [] self.mac.init_hidden(batch.batch_size) for t in range(batch.max_seq_length - 1): # -------------------------------------------------------------------------------------# # NOTE: We hard-coded the forward pass arguments for experiment, we will fix this later # -------------------------------------------------------------------------------------# agent_outs = self.mac.forward(batch, t=t, test_mode=True, gumbel=True) agent_entropy = multinomial_entropy(agent_outs).mean(dim=-1, keepdim=True) agent_probs = th.nn.functional.softmax(agent_outs, dim=-1) mac_out.append(agent_probs) mac_out_entropy.append(agent_entropy) mac_out = th.stack(mac_out, dim=1) # Concat over time mac_out_entropy = th.stack(mac_out_entropy, dim=1) # Mask out unavailable actions, renormalise (as in action selection) mac_out[avail_actions == 0] = 0 mac_out = mac_out/mac_out.sum(dim=-1, keepdim=True) mac_out[avail_actions == 0] = 0 mix_loss = self.critic(mac_out, batch["state"][:, :-1]) mask = mask.expand_as(mix_loss) entropy_mask = copy.deepcopy(mask) mix_loss = (mix_loss * mask).sum() / mask.sum() entropy_loss = (mac_out_entropy * entropy_mask).sum() / entropy_mask.sum() entropy_ratio = self.entropy_coef / entropy_loss.item() mix_loss = - mix_loss - entropy_ratio * entropy_loss # Optimise agents self.agent_optimiser.zero_grad() mix_loss.backward() grad_norm = th.nn.utils.clip_grad_norm_(self.agent_params, self.args.grad_norm_clip) self.agent_optimiser.step() if t_env - self.log_stats_t_agent >= self.args.learner_log_interval: self.logger.log_stat("mix_loss", mix_loss.item(), t_env) self.logger.log_stat("entropy", entropy_loss.item(), t_env) self.logger.log_stat("agent_grad_norm", grad_norm, t_env) self.log_stats_t_agent = t_env